给定在原有任务下训练好的模型,之前的研究工作主要探索:1)设计不同的评分函数(score functions)进行模型不确定性估计[1,2];2)利用辅助异常值(auxiliary outliers)对模型进行微调[2,3],来获取及提升模型分辨分布外样本的能力。 沿着以上的研究问题,在本工作中我们发现: 在获取更优的分布外分辨能力和更好的分布内...
directly solving a minimax optimization problem to maximize the AUC score cannot achieve satisfactory ...
用(Liu et al., 2020) [10] 中的energy score来做不确定性评估。 [[2020nips-Energy-based out-of-distribution detection]] 能量得分 弥补了给定数据点的 Helmholtz free energy 与其密度之间的差距。 对于多模态数据,不同模态的密度函数可以通过相应的能量函数来估计: image.png 公式13,一个样本,每个分类器...
[ICML 2023] Change is Hard: A Closer Look at Subpopulation Shift subpopbench.csail.mit.edu Topics benchmark spurious-correlations class-imbalance subpopulation icml subgroup out-of-distribution domain-generalization ood-generalization ood-robustness subpopulation-shift icml-2023 Resources Readme Lice...
[ICML 2024 Best Paper] Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution (https://arxiv.org/abs/2310.16834) - lijrjyan/Score-Entropy-Discrete-Diffusion
设计不同的评分函数(score functions)进行模型不确定性估计 [1,2]; 利用辅助异常值(auxiliary outliers)对模型进行微调 [2,3],来获取及提升模型分辨分布外样本的能力。 考虑到分布外检测与原始任务间存在一定的任务目标差异(例如,考虑基础的分布内数据(Indistribution data, ID data)分类目标与识别分布外样本的目标...
挪威研究中心链接：https://arxiv.org/abs/2402.106343. Beyond Point Prediction: Score Matching-...
Distribution Tracker用于估计每个时间戳t处的标签分布w_t，而Prediction Optimizer用于基于估计的标签分布...
In stage 2, we select the last checkpoint of stage 1 and train with CLIP Score optimization on 4 8-V100-32G-GPU servers (expected to obtain ~34.0+ CLIP Score on the validation set at this stage). During the validation, the generated image will be dumped into _GEN_IMAGE_PATH_. # ...
Scenario C: Reproducing from training, i.e., reproducing from scratch? (1-2 weeks for training on 4x3090 GPU + 1-2 days on 24-score CPU) First, run training commands: seeUsage > Appendix: Training Scripts. Then, run certification commands: seeScenario B. ...